It’s essentially the pvalue hacking we see in social and biological sciences applied to machine learning field.

Once you set an evaluation metric it ceases to become a useful metric.